Looking for Guidance on a Custom Solution

[caymanwjeffers]

Is it possible to take an image of a pre-existing mel spectogram with a little bit of noise added, and then feed that to the diffusion model at far fewer de-noising steps? Or is it possible to give it an existing audio-prior as a starting point instead of relying on the raw model to do all the work?

I am aware that this would require editing of the existing implementation I am just wondering if it's possible at all or my understanding of this is flawed.

I am essentially looking for a way to create coherent variations of a sound instead of generating them from scratch. Thanks!